rsi.watch

Xiaoliang Wang et al. (12 authors) · 2026-10-13

ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution

“Across seven benchmarks, three rounds raise Qwen3.5-35B-A3B's mean question-answering accuracy from 34.57% to 56.04% and mean classification F1 from 24.27 to 34.03.”
Who
Xiaoliang Wang et al. (12 authors)
Date
2026-10-13 Version 1 was posted on 2026-10-13. The text we archived and checked is version 2, posted on 2026-10-26.
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ inconsistent results. The three-seed table starts from the same base score as S7, 30.15, but gives 37.82 at R3, against S7 46.61 in the main results, and 33.40 at R1, against 43.22. The paper does not say how the two sets of numbers relate. “Mean ECG-task scores rise from 33.40 at R1 to 37.82 at R3, with SD between 0.22 and 0.29 across trained rounds.”
how we rate evidence
Replication
No rerun found. The paper says the implementation and training-data files are not released. We found no rerun by another group. Searched up to 2026-10-28. all claims
Checked
2026-10-28, against the source page
Links
Source Paper

Also on the page

  • The model being trained is Qwen3.5-35B-A3B; its language model is fine-tuned and its vision encoder stays frozen. A data factory builds questions from measured training ECGs, or edits a waveform toward a requested feature and remeasures it first; a question is kept only if its evidence checks out. Construction failures lead to revisions of the waveform operators, prompts and skills that build the data.
  • Each round, three data recipes of 10,000 questions each are trained from the same parent and compared on a separate selection set. The paper says evaluation data are held out from generation and selection and patient-disjoint from training sources. We found no statement of the datasets the training, feedback and selection ECGs come from.
  • With operator updates frozen, reaching 10,000 accepted questions took 36,723 candidates instead of 16,193, and R1 S7 was 39.29 instead of 43.22.
  • S7 gains per round were 13.07, 3.01 and 0.38, and ECG-QA, CPSC2018 and CODE-15% dropped in R3. The authors: "In this run, continued correction on the feedback set is therefore insufficient to characterize external progress."

In the self-improving AI catalog

  • ECG-RSI: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that ECG-RSI cites, oldest first. Under each: where that work sits on the map, then where ECG-RSI sits (changes in bold), and what ECG-RSI's own paper says about it. Its whole family on the citation map

  1. 2022 STaR uses it

    Model weights · becomes the next version · improves the task agent

    “As in iterative reasoning supervision [58], explanations supply training targets; paired variants must preserve the verified image, question, options, answer, and clinical facts.”
    “Earlier self-improvement methods revise responses or retain verbal feedback [33, 39], generate instruction data [45, 49], or iteratively refresh reasoning supervision [58].”

  2. 2025 Darwin Gödel Machine uses it

    Tools, own code → Model weights · becomes the next version · improves the improver → improves the task agent

    “Feedback and training-side probes define three recipe families from the same parent \theta_{t} [62, 65]:”
    “A complementary line changes the mechanism of improvement: agent code and workflows [59, 62], skills [46], data preparation [53, 52], or environments and search strategies [65, 31].”

  3. 2026 Dream-RSI uses it

    Tools, own code → Model weights · becomes the next version · improves the task agent

    “Feedback and training-side probes define three recipe families from the same parent \theta_{t} [62, 65]:”
    “A complementary line changes the mechanism of improvement: agent code and workflows [59, 62], skills [46], data preparation [53, 52], or environments and search strategies [65, 31].”

  4. 2026 RRSI uses it

    Tools, own code → Model weights · saved for later runs → becomes the next version · improves the task agent

    “Apply the predesignated output-validity and non-target regression tolerances, then retain the highest-scoring qualifying checkpoint and its data [48].”
    “These updates require assessment: revision selection [63, 48], evolving self-evaluation [29], hidden evaluations [43], and pairwise candidate ranking [10] provide different ways to compare candidates.”

  5. 2026 ECG-RSI

    Model weights · becomes the next version · improves the task agent

Also cites Reflexion 2023, Self-Refine 2023, STOP 2023, Hyperagents 2026, Weco AIDE² 2026

← Ledger Report an error